A Data Quality-aware Deep Autoencoder with Quantum-inspired Feature Encoding for DNS Anomaly Detection
Year: 2026
Type: article
Abstract: Traditional Intrusion Detection Systems (IDSs) are difficult to use against DNS-related attacks such as tunnelling, data exfiltration and command-and-control communication, because DNS traffic is high-dimensional, imbalanced and noisy.This paper proposes a framework for DNS anomaly detection based on deep representation learning, latent-space analysis and quantum-inspired feature encoding.A deep autoencoder compresses DNS traffic features into a compact latent representation; K-Means clustering gives an exploratory view of the learned latent space; each latent feature is projected onto the unit circle through a π-scaled angular transformation, generating cosine-sine representations; and a supervised classifier performs binary detection.The framework is assessed on two benchmarks under a strict leakage-controlled protocol.On CIC-Bell-DNS-EXF-2021 it achieves 79.25% accuracy, 98.82% attack recall and consistent 5-fold cross-validated performance (79.65% ± 0.35%).A data-quality audit shows that 56.8% of its rows are exact duplicates and 65.3% are label-conflicting, which identifies duplicate leakage as one source of inflation in the accuracies above 99% previously reported on it; a controlled group-disjoint re-evaluation attributes 1.94 accuracy points to that leakage directly, so it does not by itself account for the whole distance between those figures and the deterministic upper bound of 85.63% that the label conflicts impose.On the clean, large-scale BCCC-CIC-Bell-DNS-2024 corpus (499,467 flows), the proposed 64-dimensional representation achieves 88.00% accuracy, 0.68 percentage points below the best raw-feature baseline, with differences of 0.0099 in AUC-ROC and 0.0178 in attack F1, and a 20-trial Bayesian search did not surpass the selected configuration.The benefit of the quantum-inspired encoding is therefore not accuracy, but that it is bounded, fixed in scale and nearly lossless (20 raw features to 16 latent dimensions, expanded to 32; 100 to 64), and lighter than the raw-feature baseline in representation size and in the resident memory of the fitted classifier, though not in training or inference time.
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Funder: Mustansiriyah University
OpenAlex ID: W7213670290